LA-ShuffleNet: A Strong Convolutional Neural Network for Edge Computing Devices

ShuffleNetV2 is a prominent player in the field of lightweight networks and has significant implications for the development of lightweight networks and edge computing. However, it has limitations, as its accuracy falls short compared to other larger models, and it is not friendly to datasets. It loses its advantage for small-sized images with fewer channels. In this study, we analysed the design structure of ShuffleNetV2 and found room for improvement in its computational complexity and accuracy. To further improve its performance, we first upgraded the ShuffleNetV2 network structure based on the lightweight network design criteria and constructed a new network model. Second, we introduced a novel attention module named the Adaptive Pooling Attention Module (APAM) and integrated it with the new network model, constructing a high-performance model referred to as LA-ShuffleNet. Then, we proposed a convolution operation acceleration strategy called Pack. Finally, we combined the two and conducted corresponding tests on the Windows and JETSON platforms. Extensive experiments indicate that our proposed model not only exhibits substantial improvements over the baseline model but also achieves noteworthy enhancements on the ImageNet dataset, with a rise of 1.4% in Top-1 accuracy and 3.6% in Top-5 accuracy, coupled with a reduction of 0.7M parameters. Moreover, its performance surpasses that of certain prevalent lightweight networks, such as MobileNet.